FR2814558A1 - METHOD AND DEVICE FOR TRAFFIC PREDICTION WITH A NEURON NETWORK - Google Patents

Abstract

The invention proposes a method and a device for predicting, on successive prediction windows (Tp), a traffic (5) on a telecommunications network, the device comprising: - characterization means (10, 11, 12 ) past traffic continuously producing information (Utau, i) characterizing the traffic passed on respective observation windows (T1); means of production (100) containing a neural network (110), and producing, at instants (to, to + Tp, to + 2, Tp, ...), information (Dj) characterizing the traffic on a next prediction window (Tp), from a series of said values (Utau, i ) characterizing the past traffic; learning means (200) driving the neural network (110) continuously.

Description

METHOD AND DEVICE FOR PREDICTING TRAFFIC WITH A

NEURONES NETWORK

  The present invention relates to a method and a device for predicting traffic with a neural network. The term traffic means here a flow of

digital data.

  It relates to the field of telecommunications networks and finds applications with dynamic resource allocation methods and devices, including but not limited to broadband networks such as Broadband Integrated Services Digital (BISDN) networks. Networks "which means integrated digital networks of

broadband services).

  Such networks offer a data transfer service to different applications whose bandwidth and quality of service (QoS) needs are not necessarily the same. This is why the properties in terms of bit rate and QoS of a communication between two applications on such a network are generally the subject of negotiation between the network and the network.

  The source application. This negotiation takes place at the opening of the communication.

  It leads to a "traffic contract". The needs of a given application may also vary during the communication, due in particular to fluctuations in the traffic of the communication. This is why it is known to apply renegotiation procedures during the communication of the parameters of the traffic contract. This allows optimal use of network resources. This renegotiation can be done at

  the request from the source application, or periodically.

  A traffic contract of a communication comprises two elements.

  On the one hand, a traffic descriptor with parameters such as the maximum data rate transmitted by the source application and the tolerance of the delay variation between the transmission of the data units (for example ATM cells or IP datagrams). ) still called jitter tolerance. And secondly a class of QoS defined for example by a parameter such as the rate of data loss. Under the contract of subscribed traffic, the network guarantees the quality of service negotiated subject of course that the source traffic respects the parameters of the descriptor. Traffic contract parameters are used for booking or resource allocation

  network requirements (size of buffers, etc ...).

  The invention can thus be applied in connection with a renegotiation technique during communication of the parameters of a traffic contract of a communication over a telecommunications network, such as an ATM network (the ATM mode is a connection oriented) or an IP network with a signaling protocol (such as RSVP) for managing communications over that IP network. Such a technique is implemented, for example, at the level of the user interface interface (UNI) with a Local Area Network (LAN). It can also be implemented at a Network Node Interface (NNI) for the interconnection of a LAN with a Wide Area Network (WAN). ), or for

the interconnection of two WANs.

  But the applications of the invention are not limited to renegotiation techniques, during communication, the parameters of a traffic contract of a communication between two applications established on a telecommunications network. Indeed, the invention applies generally to dynamic resource allocation techniques for any network element, for example a buffer memory ("Buffer" in English), which is traversed by a determined traffic ( which can regroup the traffic of several

  respective communications established through this network element).

  Resource allocation is called dynamic because it can be modified to

  adapt to traffic behavior.

  It is especially for variable bit rate ("Bit Rate") traffic, such as the traffic observed at the interconnection of local networks, that the modification of the allocated resources allows a more efficient use of the resources of the network. The traffic period is the time interval between two successive changes. It is necessary to have a method for predicting the traffic in question by the next period

traffic.

  Methods for predicting traffic with a neural network have been published in the publication "Using Neural Networks for Predicting Transient Leaky Bucket Characterizations of Data Traffic." Application to Dynamic Resource Allocation in ATM Networks, CLEROT F, GOUZIEN P, FERAUD R, GRAVEY A, COLLOBERT D, ATM'98, IFIP Workshop on Performance Modeling and Evaluation of ATM Networks, and the publication "Dynamical resource reservation scheme in an ATM network using neural network-based traffic prediction" CLEROT F, GOUZIEN P, BENGIO S, GRAVEY A, COLLOBERT OF ATM'97, IFIP Workshop on Performance Modeling and

Evaluation of ATM Networks.

  In the patent application FR-A-2,745,455 of the same applicant, it is further disclosed a method of renegotiating, during communication, the parameters of a traffic contract of a communication over an ATM network, which comprises a step of predicting, for a predetermined period of time between two successive renegotiations, the maximum length that would reach a virtual queue fed by the communication traffic for a given output rate of said queue. This prediction is made using a neural network based on information characterizing the traffic in a network.

  period preceding the moment of renegotiation.

  A neural network is defined by a structure and parameters.

  The structure of a neural network refers to the number of layers of units it contains, and to the number of units of each of these layers. At least, a neural network comprises an input layer and an output layer, between which one or more hidden layers may optionally be provided. If the unit layers are ordered from the input layer to the output layer, the units of a given layer are connected to all the units of the previous layer (except for the layer of input) and all units of the next layer (except for the output layer) by synapses. Each synapse corresponds to a parameter of the neural network, also called weight. The value delivered by each unit (other than the input layer units) is the value of the units of the previous layer weighted by the weights

  correspondents. It should be noted that weights may possibly be zero.

  A neural network is particularly suitable for predicting the behavior of a process, such as the traffic of a communication over a telecommunications network, from the observation of this process on a window of the past. During a learning phase, the neural network is trained for this purpose. The learning consists, from a given structure of the network, to optimize the set of weight of the network and to select the one of these sets of weights which gives the best performances in term of precision of the prediction. During learning, the network of

  Neurons are fed into input by a learning sequence.

  In the prior art illustrated by the publications and the above patent application, the training sequence is a set of learning data considered to exhaustively represent the possible traffic behaviors. These data are determined during a measurement campaign carried out at the location of the telecommunications network where it is desired to implement the prediction. Learning is done before installation and

  commissioning of the prediction means, ie "offline".

  It is therefore necessary to carry out a measurement campaign and a specific learning for each place of the network where it is desired to install means for predicting traffic. In addition, these prediction means can not adapt to the subsequent evolution of traffic behavior, which may affect their performance. These means can not therefore be installed in the long term, without planning new

  measurements and learning of the corresponding neural network.

  The object of the invention is to overcome these disadvantages of the prior art. This object is achieved, in accordance with the invention, by means of a prediction method, on successive prediction windows, of a traffic on a telecommunications network, according to which: a) information is permanently generated characterizing past traffic; on the respective first observation windows; b) at determined times, information is generated characterizing the future traffic on a next prediction window, using prediction means comprising a neural network, and from a series of said information characterizing the past traffic;

  c) the neural network is permanently trained.

  In other words, the invention provides for "on-line" learning of the neural network. Information characterizing past traffic is continuously generated and the weights of the neural network can be dynamically modified, based on this information, to adapt to changes in traffic behavior. The resources of the network of

  telecommunications can then be exploited optimally.

  For convenience, the information produced in steps a) is sometimes called observed values and the information produced in step b) is

  sometimes called predicted values in the following.

  Information characterizing the traffic on a determined characterization window (observation window for the past traffic or prediction window for the future traffic) is preferably constituted by a higher approximation of a point of the traffic arrival curve. on said characterization window, ie by an estimation of the maximum quantity of data received during a time interval of determined duration on said characterization window. The predicted value is then homogeneous at a rate, expressed in Mbits / s (megabits per second). Other information could also be used to characterize the traffic, such as the average time between the receipt of two consecutive data units (ATM cells or IP datagrams). In any case, we do not try to predict the exact value of the traffic on the next prediction window, but the value

  a traffic envelope on this window.

  The permanent learning of the neural network also makes it possible to dynamically modify the target prediction error rate. This rate is for example expressed as a percentage of erroneous predictions. A prediction is said to be erroneous when the value characterizing the future traffic on the considered prediction window that is produced subsequently turns out to be less than a value characterizing the traffic that has actually been observed on said window. The forecast prediction error rate makes it possible to adjust the pessimism of the prediction. Indeed, the lower the rate, the higher the predicted values. The invention also proposes a device for implementing the method. This device comprises a prediction device, on windows of successive predictions, of a traffic on a telecommunications network, comprising: - means of characterization of past traffic continuously producing information characterizing the traffic passed on windows of observation respectively; means of production containing a network of neurons, and producing, at moments, information characterizing the traffic on a next prediction window, from a series of said values characterizing the past traffic; learning means driving the neural network permanently. Other features and advantages of the invention will become apparent

  on reading the following description. This is purely illustrative and

  should be read with reference to the accompanying drawings, in which: - in Figure 1: a graph showing the data rate of a communication over a communication network as a function of time; in FIG. 2: the symbolic diagram of a device for implementing the method according to the invention; in FIG. 3: the symbolic diagram of a device for implementing the method according to a preferred embodiment; in FIG. 4: a graph showing the performance of the prediction means as a function of the target prediction error rate; - Figure 5: a graph showing the error rate obtained according to the target error rate; in FIG. 6: the diagram of an exemplary application of a device

according to the invention.

  In the graph of FIG. 1, the curve indicates, as a function of time t, the bit rate B (t) of the data flow corresponding to the traffic of a given communication over a telecommunications network. Such a curve is called "trace" in the jargon of the skilled person. It gives a graphical representation of the traffic of the communication. In Figure 1, there is a moment at which renegotiation of the online traffic contract of the communication is carried out. We denote Tp the time interval between two successive renegotiations. In general, the duration of the time interval Tp is constant, renegotiations being carried out periodically, at times to, to + Tp, to + 2.tp, ... regularly spaced in time. This is called the traffic period to designate the time interval between two negotiations. The traffic prediction method according to the invention aims at estimating, in a particular node of the telecommunication network and at times to, to + Tp, to + 2.Tp, etc., the behavior of future traffic over time windows. , called prediction windows corresponding preferably to the traffic periods Tp, from the observation of the behavior of the traffic passed on observation windows T1. The width of the next traffic period or prediction window Tp corresponds to the prediction horizon. The observation windows T1 are consecutive time intervals, which divide the time spent, that is, prior to the instant considered to, to + Tp, to + 2.Tp, etc. It will be noted that, in the general case, the observation windows T1 do not necessarily coincide with the prediction windows

Previous Tp.

  The behavior of the traffic can be characterized by information such as a higher approximation of a value of the traffic arrival curve over a determined time interval called characterization window. For past traffic, this time interval is the observation window T1. For future traffic, this time interval is the prediction window Tp. Thus, at each instant to, to + Tp, to + 2Tp, etc., a higher approximation of a value of the traffic arrival curve (future) is produced on the corresponding prediction window Tp, starting from a series of values of the arrival curve of the

  traffic passed on respective T1 observation windows.

  In what follows, we first recall the theoretical notion of arrival curve and describe a past traffic characterization method that allows, in practice, to produce a higher approximation of a value of the curve of arrival of past traffic on a characterization window

  Any determined, with some approximations.

  From a theoretical point of view, and for a given traffic, a value U (t) of the traffic arrival curve on a determined TO characterization window, is defined by: U (T) = max 0 <t < To-t (.t + tB (u} of) (1)

  o B (u) is the instantaneous rate of traffic at instant u.

  The value U (t) of the arrival curve of the traffic on the window To is therefore the maximum amount of data received during a time interval of duration t, less than To, on the characterization window To. arrival can be used for the dimensioning of the resources of the telecommunications network (size of memories, flow of links etc ...) according to criteria of quality of service (rate of losses, time of delivery, etc ...). The complete characterization of the traffic on the window To consists of the whole arrival curve, ie of all the values U (t) for t understood.

between O and To (0 <t <To).

  However, in an exemplary implementation of the method according to the invention, the past traffic is characterized, over a determined characterization interval To, only by certain values of U (T). In the following, we describe the operation of a device for producing a single value U (-c), for a determined value X. If one wants to produce several values U (t1), U (2), .... U (tq) for q values 1l, 2, ..., Tq determined, it is enough to put in parallel q such devices, dedicated each one to the production of one of these values

U (2), .., U (: q).

  In fact, a method of on-line characterization of past traffic is explained below which makes it possible, in practice, to generate an approximation online.

higher than the value U (t).

  Let us suppose that the time is divided into time intervals of duration e much smaller than a, and that it is possible to measure in line the quantity Bj of data arrived at each of these time intervals. By denoting E (x) the integer function of x, we have the following approximation: jt + tB (u) = zj = E (t + c) / OBj (2): -j = E (t / 0 e More precisely, by setting k = E (t / 0) and framing a time interval of duration t in the time intervals of multiple durations of 0, we obtain: kj = k + E (-t / e), .,. <my max k (, j = Ek (3) Bj) <U (,) <max k (zj = _k (2Bj) (3)

with 0 <k <E ((To-t) / O).

  These maxima are easily obtained from the quantities Bj measured online. The approximation is obviously better if 0 is small in front of c, but the computation load increases in inverse proportion of 0. In an example X is equal to 1 second, and 0 is equal to 0.1 second. Another

  approximation consists in retaining only the right-hand side of the inequality (3).

  This gives a higher approximation of the maximum amount of data of the communication traffic which has been received at the considered node of the telecommunications network during a time interval t

  on the characterization window To.

  The characterization method described above is implemented by a device comprising a characterization module co-operating with a measurement module of the data quantities Bj arriving on respective time intervals 0. These modules are advantageously made in the form of software. FIG. 2 diagrammatically shows a device for setting

  implementation of the prediction method according to the invention.

  The traffic of the communication is symbolized by the arrow 5. A measurement module 7 produces the quantities Bj on-line A first characterization module 11 permanently produces UC values, i characterizing the traffic passed on consecutive observation windows T1 . This characterization module 11 implements the on-line characterization method described above, based on the quantities Bj supplied to it by the module of

measure 7.

  The UTi values are provided at the input of prediction means 100 comprising a neuron network 110. These means 100 produce a predicted value Di of the traffic arrival curve 5 on a determined prediction window, from a series of last Uji values. All Usi values characterizing past traffic are therefore not taken into account, but only the most recent ones. These values U ,, i at the input of the units of the input layer of the neural network 110, whose output layer comprises only one unit, which delivers the value Di. This value is homogeneous at a rate. It corresponds to the maximum value of the traffic flow

  of the communication on the next Tp prediction window.

  The neural network 110 includes a learning cell.

  This includes an "apprentice" neuron network 110Oa of the same structure as the neural network 110. The neural network 110 is permanently driven. For this purpose, the device comprises a learning module 200, which permanently drives the neural network from the series of observed values U i. This continuous learning actually affects the "apprentice" neural network 110a of the learning cell. Indeed, the parameters of the neural network 110 are

  frozen so that they are not modified by learning as such.

  However, the learning module 200 continuously compares the performance of the neural network 110 to that of the "apprentice" neuron network 110a, the parameters of the "apprentice" neuron network 110a being copied back into the neural network 110, under the command of the learning module 200, when the performances of the first are better than

those of the second.

  To enable learning, the values Ui delivered by the characterization module 11 are saved in a first database 21. From there, they are then provided as input to the "apprentice" neuron network 110a for learning. The learning consists in optimizing the weight play wm of the "apprentice" neuron network 110a by producing an i value characterizing the traffic on the next prediction window Tp using the "apprentice" neuron network 110 a, and comparing this value D'i to a corresponding value D "i provided by" an oracle. "This comparison is done via a" cost function "whose form is detailed below, the module 200 then determines the performance of the network. Apprentice neurons 110a by comparing the values of i and d "i. The module 200 also determines the performance of the neural network 110 by comparing the value Di delivered by the latter for the same prediction window Tp, and the

corresponding D i value.

  An "oracle" has the function of delivering values D "i corresponding to the predicted values Di or D'i The values D" i correspond to the values Di or D'i in the sense that they are comparable to them, as characterizing the traffic on

same time windows.

  In an embodiment according to the device of FIG. 2, an "oracle" is produced in the following manner. The device comprises a second characterization module 12 which generates a series of values U "i characterizing the traffic 5 on second observation windows T2.

  also receives the quantities Bj delivered by the measurement module module 7.

  Like the first module 11, the second module 12 implements the characterization method described above. The values U ", i are saved in a second database 22. For the sake of clarity, the values read in the database 22 are denoted D" i although they correspond to the values U "i. T2 observation coincide with the prediction windows Tp earlier .The database 22 therefore behaves like an "oracle", in that it contains values characterizing past traffic actually observed, which are comparable to the corresponding predicted Di predicted values Di by the neural network 110

  and the predicted values output the prediction means 100a.

  In Figure 3, there is shown schematically a device for implementing the method according to the invention according to a preferred embodiment. In this embodiment, the observation windows T1 coincide with the prediction windows Tp passed. This simplifies the method and the device according to the invention because the information U-, i characterizing the traffic 5 on the respective observation windows T1 can directly feed the database 22 of the oracle. Indeed, they can later be compared to the predicted values Di of the maximum traffic rate 5 on the corresponding prediction windows, for the online learning of the neural network. The functions of the characterization modules 11 and 12 of the device of FIG. 2 are therefore ensured by a single characterization module 10 of the same nature as the latter. This module therefore implements the aforementioned characterization method for producing the information UT, i. These are saved in the database 22. For the sake of clarity, the information read in the database 22 is always denoted D "i,

  although they correspond to the values U, i.

  The embodiment of FIG. 3 is further distinguished from that of FIG. 2, in that the prediction means 100 comprise, in addition to the neural network 110 and its learning cell with the "apprentice" neural network 110a. a battery 120 of n prediction filters F1 to Fn having different filtering windows (filter depth). These filters have the function of producing intermediate information characterizing the traffic 5 on the next prediction window Tp from a series of information U. i characterizing the traffic passed on respective observation windows T1. For this purpose, the filters F1 to Fn are each fed by the information U, i, which are produced and delivered by the characterization module 10. The vector Vi composed of the output values DF1i to Dfni respectively of the filters F1 to Fn at the instants determined to, to + Tp, to + 2.p, ... is transmitted to the input of the neural network 110, which then produces the value Di from these values. In addition, the vector Vi is saved in the database 21 for learning

online neural network.

  In one example, the prediction horizon Tp is set to 10 seconds so that the width of the windows Tp, and therefore that of the observation windows T1, are equal to 10 seconds. In addition, the filter bank 120 comprises five filters (n = 5) F1 to F5 whose respective filtering depths are respectively 10, 30, 50, 80 and 120 times the width of the observation windows T1. In this way, a series of ten consecutive values U i enter the filter window of the filter F1, a series of thirty values U i

  enter the filter window of the filter F2, etc ...

  The function of the filter bank 120 is to pre-process the information U ,, i characterizing the past traffic. This pre-treatment allows the prediction means 100 to offer satisfactory performance, while including a network of neurons 110 of fairly simple structure. The simpler the structure of the neural network 110, the less weight it must have, the values of which must be determined by learning the neural network, and therefore the learning is faster. The pre-treatment carried out by the filter bank 120 thus makes it possible to use a network of neurons with a fairly simple structure, the learning of which is fast enough to be carried out continuously, without sacrificing the performance of the means of communication.

  prediction in terms of the precision of the prediction. In other words, the

  treatment by the battery of predictive filters makes it possible to resolve the compromise between high performances of the prediction means (which assume, without the pre-treatment proposed by the invention, a network of neurons of relatively complex structure, of which with a lot of weight) and speed of learning the neural network (which assumes a network of

  neurons of relatively simple structure, with relatively little weight).

  Tests have shown that, in the aforementioned example of a filter bank 120 comprising five filters, a neural network consisting of a single perceptron (neuron network with an input layer and an output layer, without any hidden layer) is sufficient to achieve satisfactory performance. In this example, the input layer of the neural network comprises six units: five inputs connected respectively to the output of the five filters F1 to Fn to receive the intermediate values respectively DF1i to DF5i, and a sixth input for the bias of the network of neurons. In addition, the output layer has only one unit, which delivers the value Di. Thus, the neural network has only six weights Wm, o m

  is an index between 1 and 6. So it is actually quick to train.

  Other perceptron architectures have also been tested with a hidden layer, which comprises two to five units, without noticeable improvement in performance in terms of accuracy of the prediction. That is why we prefer the use of a simple perception, whose learning is

the fastest.

  Preferably, the predictor filters F1 to Fn are Kalman filters having different respective filter depths. Kalman filters are indeed simple to implement. In addition, they have a low computational cost. In addition, they make it possible to effectively continue a possibly non-stationary process as is the case, for example, for the traffic of a communication over a telecommunications network. However, other types of prediction filters may be used, for example a filter operating a simple arithmetic mean of the series of information entering the filter filter window, or a filter performing an exponential average of

these values.

  The implementation of a Kalman filter requires to make a hypothesis on the relation between the information of the space of entry and those of the space of exit. This relation is linear in the case of the classical Kalman filter, and nonlinear in the case of the extended Kalman filter ("Extended Kalman Filter"). In the example, the information of the input space of the prediction filters F1, F, F3, F4 and F5 are constituted by the series of respectively 10, 30, 50, 80 and 120 UTi information. It is assumed that these series are realizations of a respective Gaussian random variable, whose mean pm and the variance am, o m is an index between 1 and n, are estimated in line by the respective Kalman filters. In one example, an information DF1i to DFni at the output of the Kalman filter respectively F1 to Fn, is then given by: DFmi = Pm (4) Thus, the information DFmi is the Gaussian average! Tm of the information U., i which enter the filter window of the filter Fm considered. More concretely, this is an estimate of the maximum traffic flow on the

  next Tp prediction window.

  If it were possible to determine an optimum filtering window size, the prediction means 100 could, if need be, include only one Kalman filter, without a neural network. Nevertheless, this is not

  not the case, as shown in curves C1 to C5 of the graph in Figure 4.

  For a given trace, these curves represent, as a function of the rate of prediction errors τ referred to (expressed as a percentage of erroneous predictions), the values of the maximum traffic rate (expressed in Mbps), as predicted by the filters. prediction respectively F1 to F5. However, we see that the best prediction (that is, the predicted value of the lowest maximum traffic rate) is not obtained with the same prediction filter, depending on the value of s considered. Analyzes show that, for a given value of c, the best prediction is not obtained with the same Kalman filter according to the trace considered, which is why a Kalman filter can not be the only means of prediction. These remarks are not specific to Kalman filters but apply to all prediction filters, because it is a problem related to

  depth of filter prediction filters, whatever the type.

  Therefore, in the embodiment of Figure 3, the neural network 110 has the function of combining the intermediate values DFmi characterizing the traffic on the next prediction window, for m between 1 and n. The curve C0 of FIG. 4 represents, for said determined trace, as a function of the prediction error rate c targeted, the values of the maximum traffic rate, as predicted by the neural network 110 from the intermediate values DF1i to DF5i which are produced by the prediction filters respectively F1 to F5. As can be seen, whatever the value of c considered, the value of the curve C0 is lower than that corresponding to the curves C1 to C5. The performance of the neural network thus fed in by the intermediate values produced by the filters F1 to F5 are better than those of each of the filters F1 to F5 taken individually. The permanent learning of the neural network 110 makes it possible to give a significant weight only to the one or those of these intermediate values which are the best according to the circumstances related to the traffic. In other words, the permanent learning of the neural network 110 makes it possible to retain and combine only the prediction filter or filters F1 to Fn are the

  more appropriate, depending on the circumstances related to the traffic.

  A learning base is composed of pairs of values (Ut, i, D "i) in the example of FIG. 2, or pairs of values (Vj, D" i) in the example of FIG. 3, these pairs of values being called learning patterns in the jargon of those skilled in the art and in the following. In each learning pattern, the first value corresponds to the second value, in the sense indicated above. These values are all values characterizing the traffic on time windows included within a window of the past considered for learning. The first values of each learning pattern are read from the database 21 and the seconds are read from the database 22. The structure of the databases makes it possible to establish the correspondence between the first and the second

  value of each learning pattern.

  Learning is based on the so-called technique of "minimizing structural risk" in the jargon of those skilled in the art. It consists in separating the learning base into two distinct sets, for example by random draw: a learning set which serves to drive the "apprentice" neural network 110a, this learning being done by applying the learning patterns and optimizing the weight play to achieve the best performance, with the understanding that each pattern can be applied multiple times, until the learning stops under the conditions below; a validation set which is used to evaluate the performance of the "apprentice" neuron network 110a, and the performance of the predictor neuron network 110, the apprenticeship of the "apprentice" neuron network stopping when its performance is degraded on the validation set, and the weights of the predictor neuron network 110 being replaced by those of the "apprentice" neuron network 110a when the performance of the second

  are better than those of the first on this same validation set.

  Further information on the so-called "structural risk minimization" technique can be found in V. VAPNIK, "The

  Nature of Statistical Learning Theory ", Springer-Verlag, Berlin.

  In addition, a preprocessing is applied to the learning patterns (UTi, D "i) or (Vi, D" i) so as to normalize them. This normalization uses a normalization function F applied to the values Uc, i or Vi and to a normalization function G applied to the values D "I. These functions may be identical but this is not necessarily the case. , they are defined according to a variable x, by: F (x) = Log10 (x) (5) xmax / l0 G (x) 2x x -1 (6) xmax

  O xmax is the maximum value that the variable x can take.

  In the context of the invention, xmax is the maximum value that can be

  take the traffic flow. This value depends of course on the application.

  In the example, of the traffic of a communication on a telecommunication network, a typical value of the maximum bit rate is currently equal to 10 Mbits / s. It can be seen that the function G above makes it possible to normalize the values of D "i between -1 and +1, which guarantees a good definition of the

  proposed cost function (see below).

  Preferably, the values UTi or Vi saved in the database 21 values thus standardized. Likewise, values D "i saved in the database 22 are the values thus standardized.

  Learning pattern consists of standardized values.

  The learning module is for example made in the form of software. In one example, the learning module 200 implements a standard algorithm for backpropagating the error with, for cost function, a function containing the aforementioned parameter ca, which determines the prediction error rate s targeted . This function of cost F is for example the following: F = (D "i-D'i) 2 + ox.Log [1+ (1 + D'i) / (1 + D" i)] (7) The first term represents the standard quadratic cost (D "i-D'i) 2. The second term contains a parameter a whose value can be set according to a precision objective of the prediction expressed in terms of error rate This parameter is preferably dynamically modifiable, depending on the error rate targeted S. For this purpose, the target error rate is delivered to the learning means 200 as a control parameter. adjust the pessimism of the prediction to obtain a determined real error rate The value of the parameter has to be used for learning the neural network in order to obtain the rate

  s prediction error can be determined by the following method.

  This method is implemented by the learning module 200.

  Suppose the current value of the parameter is equal to ao. In a sufficiently large window of the past, the values of the report are calculated

  Ri = (11 + D'i) / (11 + D "i) on all the learning patterns of this window.

  These values are then ordered in ascending order, the lowest corresponding to the learning patterns for which the prediction by the "apprentice" neuron network 110Oa was the riskiest, in that the difference between the characteristic value of the traffic that was actually observed on a given time window of the past and the value that was predicted by the network 110a for this time window was the smallest (or even negative). The determined learning pattern is then selected which corresponds to the predicted error rate s, and a variable R is given the value of the corresponding ratio Ri. For example, if one has a thousand learning reasons, the determined learning pattern would be, for an error rate E referred equal to 1%, the tenth of the ordered list. Determining the value of the parameter a which makes it possible to obtain the prediction error rate E referred to, then amounts to looking for the value of a, different from aco, which for said determined learning pattern would have given the value I to the variable R. This amounts to saying that a is given by: ao = aO.Log (1 + R) / Log (2) (8) On the graph of FIG. 5, the curve shows, for a given trace, the rate real prediction error is obtained with the method of determining the parameter ac described above, according to the target prediction error rate ú. As can be seen from this graph, the real rate obtained is almost equal to the target rate, whatever the value of it. This shows the effectiveness of the above method of adjusting the prediction error rate E referred to using the parameter a. The value of the ct parameter can be dynamically changed (i.e., online), but there is no guarantee regarding the performance of the prediction means during a transitional period

following this change.

  As will be understood, the characteristics of the device that determine the width of the prediction windows (ie, the prediction horizon) and the width of the observation window or windows T1 and / or T2, do not appear. not at the level of the neural network 110. These parameters of the device are in fact determined by the characteristics of the characterization modules 11 and / or 12, the filters F1 to Fn, and the size of the storage means constituted by the databases. . They can also be dynamically modified (iLe., Online), without any guarantee regarding the performance of the means of

  prediction during a transitional period following this modification.

  In the description of a preferred embodiment of the invention which

  precedes, we have chosen as a value characterizing the traffic passed on a first observation window T1 or T2 which is a higher approximation of the quantity of data received during a time interval r of determined duration on said observation window T1 or T2, produced from data quantities Bj received over respective determined time intervals 0. Preferably, the ratio between the width of the observation window T1 or T2 and the duration of the time interval T- is between 10 and 100. Such values make it possible to obtain a characterization of past traffic with precision sufficient, without implying computation times incompatible with the application. Similarly, the ratio between the duration of the time interval -r and that of the time interval 0 is between 10 and 100. Such values make it possible to obtain with a sufficient accuracy, respectively a characterization of past traffic. and an approximation of the value of the arrival curve of the traffic passed with sufficient accuracy, without involving

  computation time incompatible with the application.

  In Figure 6, there is shown an example of application of a device according to the invention. A traffic prediction device 1 according to the diagram of FIG. 2 or FIG. 3 is included in a network node interface (NNI) 2 for the interconnection of a local area network 3 (ILAN) to a network of networks.

  telecommunications 4 such as an ATM network (or an IP network).

Claims (23)

  A method of prediction, on successive prediction windows (Tp), of a traffic (5) on a telecommunications network, according to which: a) information (Uni) is continuously generated characterizing the traffic passed on first respective viewing windows (T1); b) at determined instants (to, to + Tp, to + 2.Tp, ....), information (Di) is generated characterizing the traffic on a next prediction window (Tp), using prediction means (100) comprising a neural network (110), and from a series of said information (UTi) characterizing past traffic;   c) driving the neural network (110) continuously.   2. Method according to claim 1, wherein the information (Un, i) characterizing the traffic passed on the first observation window (T1) are   saved in a first database (21).   3. Method according to claim 2, wherein, in step c), the neural network (110) is driven from a series of said information (Un, i) characterizing the past traffic, which are read in the first base of data (21).   4. The method according to claim 1, wherein step b) consists of: b) producing in parallel, n intermediate information (Dli-Dni) characterizing the traffic on the next prediction window (Tp), using n respective filters (F1-Fn) prediction means (100), these filters (F1-Fn) having respective different filter windows and being fed each   by the series of information (UTi) characterizing past traffic.   b2) producing said information (Di) characterizing the traffic on the next prediction window (Tp) using the neural network (110) receiving as input the n sets of intermediate information (Kl i-Kni) characterizing the traffic   on the next prediction window (Tp).   5. Method according to claim 4, wherein said intermediate information (Dli-Dni) characterizing the traffic on the next window of   prediction (Tp) are saved in a first database (21).   6. Method according to claim 5, wherein, in step c), the neural network (110) is driven from respective series of said n intermediate information (D! -Dn) characterizing the traffic on the next window of prediction (Tp), which are read in the first database (21).   7. Method according to any one of claims 1 to 6, following   which furthermore produces information (U ',, i) characterizing the traffic passed on respective second observation windows (T2), which coincide with   previous prediction windows (Tp).   8. The method of claim 7, wherein said information (U'I, i) characterizing the traffic passed on said second observation windows.   (T2) respectively, are saved in a second database (22).   9. Method according to any one of claims 1 to 6, following   which said first observation windows (T1) coincide with the preceding prediction windows (Tp) and according to which said values (UT, i) characterizing the traffic passed on said first observation windows (T1)   respective ones are stored in a second database (22).   The method according to claim 8 or claim 9, wherein in step c) the neural network (110) is driven from a series of values (UTi, U ', i). ) characterizing the traffic passed on said preceding prediction windows (Tp), which are read in the second base of data (22).   11. Method according to any one of claims 4 to 10, following   which filters (F1-Fn) are Kalman filters.   12. Method according to any one of claims 4 to 11, following   which the neural network is a simple perceptron.   13. Process according to any one of the preceding claims,   according to which, in step c), a feedback algorithm is implemented.   propagation of the error with, for cost function, a function containing a   parameter (a) determining a target prediction error rate.   The method of claim 13, wherein said parameter is dynamically modifiable.   15. Process according to any one of the preceding claims,   according to which, in step c), the performances of the neural network (110) are compared with those of an "apprentice" neuron network (110 a) having the same structure as the neural network, the parameters of the network of "apprentice" neurons (110 a) being copied into the neural network (110)   when the performances of the first are better than those of the second.   16. The method according to any one of the preceding claims,   according to which a value characterizing the traffic passed on a first observation window (T1) or on a second observation window (T2) is an upper approximation of the quantity of data received during a time interval (X) of determined duration on said observation window (T1, T2), produced from data quantities (Bj) received on   respective time intervals (0).   17. The method of claim 16, wherein the ratio between the width of said observation window (T1, T2) and the duration of said interval of   time (X) is between 10 and 100.   18. A prediction device, on successive prediction windows (Tp), of a traffic (5) on a telecommunications network, comprising: - means for characterizing (10, 11, 12) past traffic continuously producing information (Ui) characterizing the traffic passed on respective observation windows (T1); means of production (100) containing a network of neurons (110), and producing, at times (to, to + Tp, to + 2.Tp,...), information (Di) characterizing the traffic on an upcoming prediction window (Tp), from a series of said values (UTi) characterizing past traffic; learning means (200) driving the neural network (110) permanently.   The device of claim 18, wherein the prediction means (100) further comprises n prediction filters (F1-Fn) having respective different filtering windows, each being designated by said information series (UTi) characterizing past traffic, and producing determined instantaneous audits (to, to + Tp, to + 2.Tp, ...) n respective intermediate information (DF1 i-Dfni) characterizing the traffic on the next window of   prediction (Tp), which are delivered at the input of the neural network (110).   20. The device of claim 19, wherein the network of   Neurons (110) is a simple perceptron.   21. Device according to claim 19 or claim 20, in   which prediction filters (F1-Fn) are Kalman filters.   22. Network node interface (2) for the interconnection of a local area network (3) to a telecommunications network (4), comprising a device (1)   according to any one of claims 18 to 21.   23, Telecommunications network (4) including an interface of   network node (2) according to claim 22.